Matthias Aßenmacher

Cem Akkus, Luyang Chu, Vladana Djakovic et al.

This book is the result of a seminar in which we reviewed multimodal approaches and attempted to create a solid overview of the field, starting with the current state-of-the-art approaches in the two subfields of Deep Learning individually. Further, modeling frameworks are discussed where one modality is transformed into the other, as well as models in which one modality is utilized to enhance representation learning for the other. To conclude the second part, architectures with a focus on handling both modalities simultaneously are introduced. Finally, we also cover other modalities as well as general-purpose multi-modal models, which are able to handle different tasks on different modalities within one unified architecture. One interesting application (Generative Art) eventually caps off this booklet.

Lukas Rauch, Matthias Aßenmacher, Denis Huseljic et al.

Deep active learning (DAL) seeks to reduce annotation costs by enabling the model to actively query instance annotations from which it expects to learn the most. Despite extensive research, there is currently no standardized evaluation protocol for transformer-based language models in the field of DAL. Diverse experimental settings lead to difficulties in comparing research and deriving recommendations for practitioners. To tackle this challenge, we propose the ActiveGLAE benchmark, a comprehensive collection of data sets and evaluation guidelines for assessing DAL. Our benchmark aims to facilitate and streamline the evaluation process of novel DAL strategies. Additionally, we provide an extensive overview of current practice in DAL with transformer-based language models. We identify three key challenges - data set selection, model training, and DAL settings - that pose difficulties in comparing query strategies. We establish baseline results through an extensive set of experiments as a reference point for evaluating future work. Based on our findings, we provide guidelines for researchers and practitioners.

Ibrahim Tolga Öztürk, Rostislav Nedelchev, Christian Heumann et al.

Recent studies have demonstrated how to assess the stereotypical bias in pre-trained English language models. In this work, we extend this branch of research in multiple different dimensions by systematically investigating (a) mono- and multilingual models of (b) different underlying architectures with respect to their bias in (c) multiple different languages. To that end, we make use of the English StereoSet data set (Nadeem et al., 2021), which we semi-automatically translate into German, French, Spanish, and Turkish. We find that it is of major importance to conduct this type of analysis in a multilingual setting, as our experiments show a much more nuanced picture as well as notable differences from the English-only analysis. The main takeaways from our analysis are that mGPT-2 (partly) shows surprising anti-stereotypical behavior across languages, English (monolingual) models exhibit the strongest bias, and the stereotypes reflected in the data set are least present in Turkish models. Finally, we release our codebase alongside the translated data sets and practical guidelines for the semi-automatic translation to encourage a further extension of our work to other languages.

Björn Deiseroth, Max Meuer, Nikolas Gritsch et al.

Large Language Models (LLMs) have reshaped natural language processing with their impressive capabilities. However, their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. This study introduces the Divergent Token Metrics (DTMs), a novel approach to assessing compressed LLMs, addressing the limitations of traditional perplexity or accuracy measures that fail to accurately reflect text generation quality. DTMs measure token divergences that allow deeper insights into the subtleties of model compression, in particular, when evaluating components' impacts individually. Utilizing the First Divergent Token Metric (FDTM) in model sparsification reveals that 25% of all attention components can be pruned beyond 90% on the Llama-2 model family, still keeping SOTA performance. For quantization, FDTM suggests that more than 80% of parameters can be naively transformed to int8 without special outlier management. These evaluations indicate the necessity of choosing appropriate compressions for parameters individually -- and that FDTM can identify those -- while standard metrics result in deteriorated outcomes.

Philipp Koch, Gilary Vera Nuñez, Esteban Garces Arias et al.

The Bavarian Academy of Sciences and Humanities aims to digitize its Medieval Latin Dictionary. This dictionary entails record cards referring to lemmas in medieval Latin, a low-resource language. A crucial step of the digitization process is the Handwritten Text Recognition (HTR) of the handwritten lemmas found on these record cards. In our work, we introduce an end-to-end pipeline, tailored to the medieval Latin dictionary, for locating, extracting, and transcribing the lemmas. We employ two state-of-the-art (SOTA) image segmentation models to prepare the initial data set for the HTR task. Furthermore, we experiment with different transformer-based models and conduct a set of experiments to explore the capabilities of different combinations of vision encoders with a GPT-2 decoder. Additionally, we also apply extensive data augmentation resulting in a highly competitive model. The best-performing setup achieved a Character Error Rate (CER) of 0.015, which is even superior to the commercial Google Cloud Vision model, and shows more stable performance.

Esteban Garces Arias, Julian Rodemann, Meimingwei Li et al.

Decoding from the output distributions of large language models to produce high-quality text is a complex challenge in language modeling. Various approaches, such as beam search, sampling with temperature, $k-$sampling, nucleus $p-$sampling, typical decoding, contrastive decoding, and contrastive search, have been proposed to address this problem, aiming to improve coherence, diversity, as well as resemblance to human-generated text. In this study, we introduce adaptive contrastive search, a novel decoding strategy extending contrastive search by incorporating an adaptive degeneration penalty, guided by the estimated uncertainty of the model at each generation step. This strategy is designed to enhance both the creativity and diversity of the language modeling process while at the same time producing coherent and high-quality generated text output. Our findings indicate performance enhancement in both aspects, across different model architectures and datasets, underscoring the effectiveness of our method in text generation tasks. Our code base, datasets, and models are publicly available.

Stefanie Urchs, Veronika Thurner, Matthias Aßenmacher et al.

With the introduction of ChatGPT, OpenAI made large language models (LLM) accessible to users with limited IT expertise. However, users with no background in natural language processing (NLP) might lack a proper understanding of LLMs. Thus the awareness of their inherent limitations, and therefore will take the systems' output at face value. In this paper, we systematically analyse prompts and the generated responses to identify possible problematic issues with a special focus on gender biases, which users need to be aware of when processing the system's output. We explore how ChatGPT reacts in English and German if prompted to answer from a female, male, or neutral perspective. In an in-depth investigation, we examine selected prompts and analyse to what extent responses differ if the system is prompted several times in an identical way. On this basis, we show that ChatGPT is indeed useful for helping non-IT users draft texts for their daily work. However, it is absolutely crucial to thoroughly check the system's responses for biases as well as for syntactic and grammatical mistakes.

Matthias Aßenmacher, Nadja Sauter, Christian Heumann

Annotating costs of large corpora are still one of the main bottlenecks in empirical social science research. On the one hand, making use of the capabilities of domain transfer allows re-using annotated data sets and trained models. On the other hand, it is not clear how well domain transfer works and how reliable the results are for transfer across different dimensions. We explore the potential of domain transfer across geographical locations, languages, time, and genre in a large-scale database of political manifestos. First, we show the strong within-domain classification performance of fine-tuned transformer models. Second, we vary the genre of the test set across the aforementioned dimensions to test for the fine-tuned models' robustness and transferability. For switching genres, we use an external corpus of transcribed speeches from New Zealand politicians while for the other three dimensions, custom splits of the Manifesto database are used. While BERT achieves the best scores in the initial experiments across modalities, DistilBERT proves to be competitive at a lower computational expense and is thus used for further experiments across time and country. The results of the additional analysis show that (Distil)BERT can be applied to future data with similar performance. Moreover, we observe (partly) notable differences between the political manifestos of different countries of origin, even if these countries share a language or a cultural background.

Luis Mayer, Christian Heumann, Matthias Aßenmacher

In recent years, large language models (LLMs) have emerged as powerful tools with potential applications in various fields, including software engineering. Within the scope of this research, we evaluate five different state-of-the-art LLMs - Bard, BingChat, ChatGPT, Llama2, and Code Llama - concerning their capabilities for text-to-code generation. In an empirical study, we feed prompts with textual descriptions of coding problems sourced from the programming website LeetCode to the models with the task of creating solutions in Python. Subsequently, the quality of the generated outputs is assessed using the testing functionalities of LeetCode. The results indicate large differences in performance between the investigated models. ChatGPT can handle these typical programming challenges by far the most effectively, surpassing even code-specialized models like Code Llama. To gain further insights, we measure the runtime as well as the memory usage of the generated outputs and compared them to the other code submissions on Leetcode. A detailed error analysis, encompassing a comparison of the differences concerning correct indentation and form of the generated code as well as an assignment of the incorrectly solved tasks to certain error categories allows us to obtain a more nuanced picture of the results and potential for improvement. The results also show a clear pattern of increasingly incorrect produced code when the models are facing a lot of context in the form of longer prompts.

Esteban Garces Arias, Hannah Blocher, Julian Rodemann et al.

Open-ended text generation has become a prominent task in natural language processing due to the rise of powerful (large) language models. However, evaluating the quality of these models and the employed decoding strategies remains challenging due to trade-offs among widely used metrics such as coherence, diversity, and perplexity. This paper addresses the specific problem of multicriteria evaluation for open-ended text generation, proposing novel methods for both relative and absolute rankings of decoding methods. Specifically, we employ benchmarking approaches based on partial orderings and present a new summary metric to balance existing automatic indicators, providing a more holistic evaluation of text generation quality. Our experiments demonstrate that the proposed approaches offer a robust way to compare decoding strategies and serve as valuable tools to guide model selection for open-ended text generation tasks. We suggest future directions for improving evaluation methodologies in text generation and make our code, datasets, and models publicly available.

Yuanhao Ding, Esteban Garces Arias, Meimingwei Li et al.

Open-ended text generation faces a critical challenge: balancing coherence with diversity in LLM outputs. While contrastive search-based decoding strategies have emerged to address this trade-off, their practical utility is often limited by hyperparameter dependence and high computational costs. We introduce GUARD, a self-adaptive decoding method that effectively balances these competing objectives through a novel "Glocal" uncertainty-driven framework. GUARD combines global entropy estimates with local entropy deviations to integrate both long-term and short-term uncertainty signals. We demonstrate that our proposed global entropy formulation effectively mitigates abrupt variations in uncertainty, such as sudden overconfidence or high entropy spikes, and provides theoretical guarantees of unbiasedness and consistency. To reduce computational overhead, we incorporate a simple yet effective token-count-based penalty into GUARD. Experimental results demonstrate that GUARD achieves a good balance between text diversity and coherence, while exhibiting substantial improvements in generation speed. In a more nuanced comparison study across different dimensions of text quality, both human and LLM evaluators validated its remarkable performance. Our code is available at https://github.com/YecanLee/GUARD.

Matthias Schöffel, Marinus Wiedner, Esteban Garces Arias et al.

Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing, yet their effectiveness in handling historical languages remains largely unexplored. This study examines the performance of open-source LLMs in part-of-speech (POS) tagging for Old Occitan, a historical language characterized by non-standardized orthography and significant diachronic variation. Through comparative analysis of two distinct corpora-hagiographical and medical texts-we evaluate how current models handle the inherent challenges of processing a low-resource historical language. Our findings demonstrate critical limitations in LLM performance when confronted with extreme orthographic and syntactic variability. We provide detailed error analysis and specific recommendations for improving model performance in historical language processing. This research advances our understanding of LLM capabilities in challenging linguistic contexts while offering practical insights for both computational linguistics and historical language studies.

Chongsheng Zhang, Shuwen Wu, Yingqi Chen et al.

Ancient manuscripts are the primary source of ancient linguistic corpora. However, many ancient manuscripts exhibit duplications due to unintentional repeated publication or deliberate forgery. The Dead Sea Scrolls, for example, include counterfeit fragments, whereas Oracle Bones (OB) contain both republished materials and fabricated specimens. Identifying ancient manuscript duplicates is of great significance for both archaeological curation and ancient history study. In this work, we design a progressive OB duplicate discovery framework that combines unsupervised low-level keypoints matching with high-level text-centric content-based matching to refine and rank the candidate OB duplicates with semantic awareness and interpretability. We compare our model with state-of-the-art content-based image retrieval and image matching methods, showing that our model yields comparable recall performance and the highest simplified mean reciprocal rank scores for both Top-5 and Top-15 retrieval results, and with significantly accelerated computation efficiency. We have discovered over 60 pairs of new OB duplicates in real-world deployment, which were missed by domain experts for decades. Code, model and real-world results are available at: https://github.com/cszhangLMU/OBD-Finder/.

Bolei Ma, Berk Yoztyurk, Anna-Carolina Haensch et al.

In recent research, large language models (LLMs) have been increasingly used to investigate public opinions. This study investigates the algorithmic fidelity of LLMs, i.e., the ability to replicate the socio-cultural context and nuanced opinions of human participants. Using open-ended survey data from the German Longitudinal Election Studies (GLES), we prompt different LLMs to generate synthetic public opinions reflective of German subpopulations by incorporating demographic features into the persona prompts. Our results show that Llama performs better than other LLMs at representing subpopulations, particularly when there is lower opinion diversity within those groups. Our findings further reveal that the LLM performs better for supporters of left-leaning parties like The Greens and The Left compared to other parties, and matches the least with the right-party AfD. Additionally, the inclusion or exclusion of specific variables in the prompts can significantly impact the models' predictions. These findings underscore the importance of aligning LLMs to more effectively model diverse public opinions while minimizing political biases and enhancing robustness in representativeness.

Stefanie Urchs, Veronika Thurner, Matthias Aßenmacher et al.

Open-access corpora are essential for advancing natural language processing (NLP) and computational social science (CSS). However, large-scale resources for German remain limited, restricting research on linguistic trends and societal issues such as gender bias. We present taz2024full, the largest publicly available corpus of German newspaper articles to date, comprising over 1.8 million texts from taz, spanning 1980 to 2024. As a demonstration of the corpus's utility for bias and discrimination research, we analyse gender representation across four decades of reporting. We find a consistent overrepresentation of men, but also a gradual shift toward more balanced coverage in recent years. Using a scalable, structured analysis pipeline, we provide a foundation for studying actor mentions, sentiment, and linguistic framing in German journalistic texts. The corpus supports a wide range of applications, from diachronic language analysis to critical media studies, and is freely available to foster inclusive and reproducible research in German-language NLP.

Enes Özeren, Matthias Aßenmacher

Chain-of-Thought (CoT) reasoning typically utilizes the discrete language space for thinking, which is inherently inefficient, as many generated tokens only enforce linguistic rules that are not required for reasoning. To bypass this, latent-space thinking allows models to think using the continuous embedding space. While existing methods for training those models show domain-specific gains, they fail to maintain performance in complex tasks, such as mathematical reasoning. We experimentally demonstrate that the Coconut approach, a form of supervised fine-tuning for latent-space thinking, is highly sensitive to design choices and exhibits several inherent limitations. To address these issues, we investigate reinforcement learning (RL) techniques -- an underexplored direction in latent-space thinking -- including GRPO and design a novel Latent RL method for directly optimizing the latent thinking steps. Our experimental results reveal that these RL-trained models still lag behind traditional language-space CoT models in the mathematical reasoning domain. We make our codebase publicly available.

Stefanie Urchs, Veronika Thurner, Matthias Aßenmacher et al.

Language corpora are the foundation of most natural language processing research, yet they often reproduce structural inequalities. One such inequality is gender discrimination in how actors are represented, which can distort analyses and perpetuate discriminatory outcomes. This paper introduces a user-centric, actor-level pipeline for detecting and mitigating gender discrimination in large-scale text corpora. By combining discourse-aware analysis with metrics for sentiment, syntactic agency, and quotation styles, our method enables both fine-grained auditing and exclusion-based balancing. Applied to the taz2024full corpus of German newspaper articles (1980-2024), the pipeline yields a more gender-balanced dataset while preserving core dynamics of the source material. Our findings show that structural asymmetries can be reduced through systematic filtering, though subtler biases in sentiment and framing remain. We release the tools and reports to support further research in discourse-based fairness auditing and equitable corpus construction.

Stefanie Urchs, Veronika Thurner, Matthias Aßenmacher et al.

Algorithmic systems such as search engines and information retrieval platforms significantly influence academic visibility and the dissemination of knowledge. Despite assumptions of neutrality, these systems can reproduce or reinforce societal biases, including those related to gender. This paper introduces and applies a bias-preserving definition of algorithmic gender fairness, which assesses whether algorithmic outputs reflect real-world gender distributions without introducing or amplifying disparities. Using a heterogeneous dataset of academic profiles from German universities and universities of applied sciences, we analyse gender differences in metadata completeness, publication retrieval in academic databases, and visibility in Google search results. While we observe no overt algorithmic discrimination, our findings reveal subtle but consistent imbalances: male professors are associated with a greater number of search results and more aligned publication records, while female professors display higher variability in digital visibility. These patterns reflect the interplay between platform algorithms, institutional curation, and individual self-presentation. Our study highlights the need for fairness evaluations that account for both technical performance and representational equality in digital systems.

Cornelia Gruber, Helen Alber, Bernd Bischl et al.

Access to high-quality labeled data remains a limiting factor in applied supervised learning. While label variation (LV), i.e., differing labels for the same instance, is common, especially in natural language processing, annotation frameworks often still rest on the assumption of a single ground truth. This overlooks human label variation (HLV), the occurrence of plausible differences in annotations, as an informative signal. Similarly, active learning (AL), a popular approach to optimizing the use of limited annotation budgets in training ML models, often relies on at least one of several simplifying assumptions, which rarely hold in practice when acknowledging HLV. In this paper, we examine foundational assumptions about truth and label nature, highlighting the need to decompose observed LV into signal (e.g., HLV) and noise (e.g., annotation error). We survey how the AL and (H)LV communities have addressed -- or neglected -- these distinctions and propose a conceptual framework for incorporating HLV throughout the AL loop, including instance selection, annotator choice, and label representation. We further discuss the integration of large language models (LLM) as annotators. Our work aims to lay a conceptual foundation for HLV-aware active learning, better reflecting the complexities of real-world annotation.

Lukas Rauch, Moritz Wirth, Denis Huseljic et al.

The advent of large language models (LLMs) capable of producing general-purpose representations lets us revisit the practicality of deep active learning (AL): By leveraging frozen LLM embeddings, we can mitigate the computational costs of iteratively fine-tuning large backbones. This study establishes a benchmark and systematically investigates the influence of LLM embedding quality on query strategies in deep AL. We employ five top-performing models from the massive text embedding benchmark (MTEB) leaderboard and two baselines for ten diverse text classification tasks. Our findings reveal key insights: First, initializing the labeled pool using diversity-based sampling synergizes with high-quality embeddings, boosting performance in early AL iterations. Second, the choice of the optimal query strategy is sensitive to embedding quality. While the computationally inexpensive Margin sampling can achieve performance spikes on specific datasets, we find that strategies like Badge exhibit greater robustness across tasks. Importantly, their effectiveness is often enhanced when paired with higher-quality embeddings. Our results emphasize the need for context-specific evaluation of AL strategies, as performance heavily depends on embedding quality and the target task.

M. Aßenmacher, P. Schulze, C. Heumann

Large Transformer-based language models are pre-trained on corpora of varying sizes, for a different number of steps and with different batch sizes. At the same time, more fundamental components, such as the pre-training objective or architectural hyperparameters, are modified. In total, it is therefore difficult to ascribe changes in performance to specific factors. Since searching the hyperparameter space over the full systems is too costly, we pre-train down-scaled versions of several popular Transformer-based architectures on a common pre-training corpus and benchmark them on a subset of the GLUE tasks (Wang et al., 2018). Specifically, we systematically compare three pre-training objectives for different shape parameters and model sizes, while also varying the number of pre-training steps and the batch size. In our experiments MLM + NSP (BERT-style) consistently outperforms MLM (RoBERTa-style) as well as the standard LM objective. Furthermore, we find that additional compute should be mainly allocated to an increased model size, while training for more steps is inefficient. Based on these observations, as a final step we attempt to scale up several systems using compound scaling (Tan and Le, 2019) adapted to Transformer-based language models.

P. Schulze, S. Wiegrebe, P. W. Thurner et al.

The objective of advanced topic modeling is not only to explore latent topical structures, but also to estimate relationships between the discovered topics and theoretically relevant metadata. Methods used to estimate such relationships must take into account that the topical structure is not directly observed, but instead being estimated itself in an unsupervised fashion, usually by common topic models. A frequently used procedure to achieve this is the method of composition, a Monte Carlo sampling technique performing multiple repeated linear regressions of sampled topic proportions on metadata covariates. In this paper, we propose two modifications of this approach: First, we substantially refine the existing implementation of the method of composition from the R package stm by replacing linear regression with the more appropriate Beta regression. Second, we provide a fundamental enhancement of the entire estimation framework by substituting the current blending of frequentist and Bayesian methods with a fully Bayesian approach. This allows for a more appropriate quantification of uncertainty. We illustrate our improved methodology by investigating relationships between Twitter posts by German parliamentarians and different metadata covariates related to their electoral districts, using the Structural Topic Model to estimate topic proportions.

M. Aßenmacher, A. Corvonato, C. Heumann

The lack of a commonly used benchmark data set (collection) such as (Super-)GLUE (Wang et al., 2018, 2019) for the evaluation of non-English pre-trained language models is a severe shortcoming of current English-centric NLP-research. It concentrates a large part of the research on English, neglecting the uncertainty when transferring conclusions found for the English language to other languages. We evaluate the performance of the German and multilingual BERT-based models currently available via the huggingface transformers library on the four tasks of the GermEval17 workshop. We compare them to pre-BERT architectures (Wojatzki et al., 2017; Schmitt et al., 2018; Attia et al., 2018) as well as to an ELMo-based architecture (Biesialska et al., 2020) and a BERT-based approach (Guhr et al., 2020). The observed improvements are put in relation to those for similar tasks and similar models (pre-BERT vs. BERT-based) for the English language in order to draw tentative conclusions about whether the observed improvements are transferable to German or potentially other related languages.

V. D. Viellieber, M. Aßenmacher

Recently it has been shown that large pre-trained language models like BERT (Devlin et al., 2018) are able to store commonsense factual knowledge captured in its pre-training corpus (Petroni et al., 2019). In our work we further evaluate this ability with respect to an application from industry creating a set of probes specifically designed to reveal technical quality issues captured as described incidents out of unstructured customer feedback in the automotive industry. After probing the out-of-the-box versions of the pre-trained models with fill-in-the-mask tasks we dynamically provide it with more knowledge via continual pre-training on the Office of Defects Investigation (ODI) Complaints data set. In our experiments the models exhibit performance regarding queries on domain-specific topics compared to when queried on factual knowledge itself, as Petroni et al. (2019) have done. For most of the evaluated architectures the correct token is predicted with a $Precision@1$ ($P@1$) of above 60\%, while for $P@5$ and $P@10$ even values of well above 80\% and up to 90\% respectively are reached. These results show the potential of using language models as a knowledge base for structured analysis of customer feedback.

Matthias Aßenmacher, Christian Heumann

Recent developments in unsupervised representation learning have successfully established the concept of transfer learning in NLP. Mainly three forces are driving the improvements in this area of research: More elaborated architectures are making better use of contextual information. Instead of simply plugging in static pre-trained representations, these are learned based on surrounding context in end-to-end trainable models with more intelligently designed language modelling objectives. Along with this, larger corpora are used as resources for pre-training large language models in a self-supervised fashion which are afterwards fine-tuned on supervised tasks. Advances in parallel computing as well as in cloud computing, made it possible to train these models with growing capacities in the same or even in shorter time than previously established models. These three developments agglomerate in new state-of-the-art (SOTA) results being revealed in a higher and higher frequency. It is not always obvious where these improvements originate from, as it is not possible to completely disentangle the contributions of the three driving forces. We set ourselves to providing a clear and concise overview on several large pre-trained language models, which achieved SOTA results in the last two years, with respect to their use of new architectures and resources. We want to clarify for the reader where the differences between the models are and we furthermore attempt to gain some insight into the single contributions of lexical/computational improvements as well as of architectural changes. We explicitly do not intend to quantify these contributions, but rather see our work as an overview in order to identify potential starting points for benchmark comparisons. Furthermore, we tentatively want to point at potential possibilities for improvement in the field of open-sourcing and reproducible research.